International Business Machines Corporation

Intel Corporation

Ayal Zaks

Shay Ben-David

Uzi Shvadron

Ira Rosen

Dorit Nuzman

Robert Valentine

Jeffrey Haskell Derby

Ulrich Weigand

Brian King Flachs

Fredy Daniel Neeser

Olga Golovanevsky

Jamie H Moreno

Alon Dayan

Mark Richard Nutter

Alexandre E Eichenberger

Thomas Winters Fox

Victor V Zyuban

Amit Golander

John Kevin Patrick O'Brien

Jesus Corbal San Adrian

Sagi Manole

Gadi Haber

Daniel Citron

David Joel Edelsohn

Nava Arela Aizikowitz

Vitaly Feldman

Nadav Levison

Shai Aharon Rubin

Arie Narkis

Revital Eres

David Joel Edelsohm

Revital Erez

Dorit Naishlos

Disclosed embodiments relate to instructions for vector operations with immediate values. In one example, a system includes a memory and a processor that includes fetch circuitry to fetch the instruction from a code storage, the instruction including an opcode, a destination identifier to specify a destination vector register, a first immediate, and a write mask identifier to specify a write mask register, the write mask register including at least one bit corresponding to each destination vector register element, the at least one bit to specify whether the destination vector register element is masked or unmasked, decode circuitry to decode the fetched instruction, and execution circuitry to execute the decoded instruction, to, use the write mask register to determine unmasked elements of the destination vector register, and, when the opcode specifies to broadcast, broadcast the first immediate to one or more unmasked vector elements of the destination vector register.

Instructions for vector operations with constant values

An apparatus and method for determining whether data needed for one or more operations is stored in a cache and scheduling the operations for execution based on the determination. For example, one embodiment of a processor comprises: a hierarchy of cache levels for caching data including at least a level 1 (L1) cache; cache occupancy determination logic to determine whether data associated with one or more subsequent operations is stored in one of the cache levels; and scheduling logic to schedule execution of the subsequent operations based on the determination of whether data associated with the subsequent operations is stored in the cache levels.

Method and apparatus for cache occupancy determination and instruction scheduling

Systems and methods are disclosed for enhancing the throughput of a processor by minimizing the number of transfers of data associated with data transfer between a register file and a memory stack. The register file used by a processor running an application is partitioned into a number of blocks. A subset of the blocks of the register file is defined in an application binary interface enabling the subset to be pre-allocated and exposed to the application binary interface. Optionally, blocks other than the subset are not exposed to the application binary interface so that the data relating to application function switch or a context switch is not transferred between the unexposed blocks and a memory stack.

Register management in an extended processor architecture

Dynamic determination of affinity between fields of structure may be determined based on accesses to the same instance. The affinity may be utilized in determining a data layout of a structure so as to optimize performance of a target program. The affinity determination may be an estimation based upon a trace of an execution of the target program. Access relation between proximate accesses to fields of the same instance may be utilized to estimate an optimized data layout of the structure.

Instance-based field affinity optimization

A compiler of a single instruction multiple data (SIMD) information handling system (IHS) identifies 'if-then-else' statements that offer opportunity for conditional branch conversion. The SIMD IHS employs a processor or processors to execute the executable program. During execution, the processor generates and updates SIMD lane mask information to track and manage the conditional branch loops of the executing program. The processor saves branch addresses and employs SIMD lane masks to identify conditional branch loops with different branch conditions than previous conditional branch loops. The processor may reduce SIMD IHS processing time during processing of compiled code of the original 'if-then-else' statements. The processor continues processing next statements inline after all SIMD lanes are complete, while providing speculative and parallel processing capability for multiple data operations of the executable program.

Management of conditional branches within a data parallel system

A method for vectorization of a block of code is provided. The method comprises receiving a first block of code as input; and converting the first block of code into at least a second block of code and a third block of code. The first block of code accesses a first set of memory addresses that are potentially misaligned. The second block of code performs conditional leaping address incrementation to selectively access a first subset of the first set of memory addresses. The third block of code accesses a second subset of the first set of memory addresses starting from an aligned memory address, simultaneously accessing multiple memory addresses at a time. No memory address belongs to both the first subset and the second subset of memory addresses.

Vectorization of program code

A data layout optimization may utilize affinity estimation between pairs of fields of a record in a computer program. The affinity estimation may be determined based on a trace of an execution and in view of actual processing entities performing each access to the fields. The disclosed subject matter may be configured to be aware of a specific architecture of a target computer having a plurality of processing entities, executing the program so as to provide an improved affinity estimation which may take into account both false sharing issues, spatial locality improvement and the like.

Architecture-aware field affinity estimation

A method for computing includes executing a program, including multiple cacheable lines of executable code, on a processor having a software-managed cache. A run-time cache management routine running on the processor is used to assemble a profile of inter-line jumps occurring in the software-managed cache while executing the program. Based on the profile, an optimized layout of the lines in the code is computed, and the lines of the program are re-ordered in accordance with the optimized layout while continuing to execute the program.

Optimization of software instruction cache by line re-ordering

A method for managing multiple values assigned to a variable during various stages of a software pipelined process executed in a computing environment. The method comprises allocating two or more slots in a vector register to two or more values associated with said variable during two or more stages of a pipeline process; and rotating values in each slot responsive to an instruction.

Software pipelining using one or more vector registers

A method for transferring data from a general purpose register (GPR) to a vector register (VR), the method including vectorially combining data in the VR from the GPR, by executing instructions of a PowerPC Instruction Set Architecture (ISA), the step of combining including splatting a low nibble from the GPR into a low nibble in each element of a first VR by executing two 'load vector for shift left' (lvsl) or 'load vector for shift right' (lvsr) and one &#x201c;vector subtract unsigned byte modulo&#x201d; (vsububm), shifting a high nibble of the GPR into a low nibble the GPR, splatting the low nibble of the GPR into a low nibble in each element of a second VR by re-executing the two lvsl or lvsr and one vsububm instructions, shifting the low nibble of the second VR into a high nibble of the second VR and combining both first and second VRs into one VR.

Splat copying GPR data to vector register elements by executing lvsr or lvsl and vector subtract instructions

A method for determining vectorization configurations in a computer processor architecture, the method including identifying a vectorizable loop in a computer program, identifying a memory access pattern of data required for implementing the loop in the architecture, computing a set of candidate configurations of resources required for vectorizing the data in the architecture, where the computing step includes configuring a vector pointer register of the architecture in support of either of reorder-on-read use and reorder-on-write use of a vector element file of the architecture, selecting one of the candidates in accordance with predefined selection criteria, and implementing the selected vectorization configuration in the architecture.

Vectorization in a SIMdD DSP architecture

An apparatus and method are provided for updating one or more pluralities of pointers (i.e. one or more vector pointers) which are used for accessing one or more pluralities of data elements (i.e. one or more vector data elements) in a multi-ported memory. A first register file holds the vector pointers, a second register file holds stride data, and a plurality of functional units combine data from the second register file with data from the first register file. The results of combining the data are transferred to the first register file and represent updated vector pointers. Furthermore, a third register file is provided for holding modulus selector data to specify the size of a circular buffer for circular addressing.

Apparatus and method for updating pointers for indirect and parallel register access

A digital signal processor (DSP) includes a SIMD-based organization wherein operations are executed on a plurality of single-instruction multiple data (SIMD) datapaths or stages connected in cascade. The functionality and data values at each stage may be different, including a different width (e.g., a different number of bits per value) in each stage. The operands and destination for data in a computational datapath are selected indirectly through vector pointer registers in a vector pointers datapath. Each vector pointer register contains a plurality of pointers into a register file of a computational datapath.

SIMD processor with concurrent operation of vector pointer datapath and vector computation datapath

A technique for improving the performance of binary tree operations involves defining an implicit binary tree structure in a memory array, and clustering of the nodes of the tree in memory in a cache-aware manner. The technique reduces memory latency by improved spatial locality of the binary tree data, and further improves cache performance through reduced size of the data objects resulting from elimination of pointers to other nodes of the tree.

Growing community of inventors

Misgav, Israel

Ayal Zaks